JP6594329B2 - System and method for facial expression - Google Patents

Description

  The present technology relates to the field of face recognition. More particularly, the present technology provides a technique for representing facial images using deep learning.

  A face recognition system is a computer application that automatically identifies or confirms a person from an image or video. Face recognition in unconstrained images is at the forefront of the algorithmic perception revolution. The social and cultural significance and benefits of facial recognition technology are extensive. However, the current performance gap in this area between the machine and the human visual system is hindering these benefits.

  The face recognition error rate when recognizing a front face in a still image taken in a consistently controlled (constrained) environment has decreased by three orders of magnitude over the past 20 years. Examples of applications include border security and smart biometrics. However, such systems can be susceptible to various factors such as lighting, facial expressions, occlusion, and aging. The performance of such systems can be significantly reduced when attempting to recognize people in unconstrained situations.

  Various embodiments of the present disclosure provide a system, method, and non-transitory configured to employ a deep neural network to align facial images, classify facial images, and verify facial images Any computer readable medium may be included. In some embodiments, a 3D-aligned face image may be generated from the 2D face image. The ID of the 2D face image can be classified based on providing the 3D aligned face image to a deep neural network (DNN). The ID of the two-dimensional face image may include a feature vector.

  In one embodiment, a set of reference points for a two-dimensional face image may be detected. The 2D face image can be used to process the 2D face image into a 3D shape to generate a 3D aligned face image.

  In one embodiment, the set of anchor points can be placed on a three-dimensional shape. Each anchor point may correspond to one reference point in the set of reference points of the two-dimensional face image. The position of each reference point in the set of reference points of the two-dimensional face image can be backprojected into a three-dimensional shape. The position of each reference point in the set of reference points of the two-dimensional face image can be projected onto a three-dimensional front image. The 2D face image can be converted to a 3D aligned face image based on the set of reference points and the set of anchor points.

  In one embodiment, the set of reference points of the 2D face image defines a set of triangles. An affine transformation is performed on each triangle in the set of triangles to transform the 2D face image into a 3D aligned face image.

  In one embodiment, the facial portion of the image may be identified by detecting a second set of reference points for the image. A two-dimensional face image may be generated by transforming the face portion of the image based on the second set of reference points.

In one embodiment, a set of anchor positions can be determined. The face portion of the image can be processed based on a set of anchor positions to generate a two-dimensional face image.
In one embodiment, it comprises a set of DNN layers, the set of layers comprising a convolutional pooling layer, a set of local coupling layers, and a set of all coupling layers.

  In one embodiment, the convolution pooling layer comprises a first convolution layer, a maximum pooling layer, and a second convolution layer. The convolutional pooling layer is configured to extract a set of features of the three-dimensionally aligned facial image.

In one embodiment, each local combined layer of the set of local combined layers is configured to extract a set of features of a 3D aligned facial image.
In one embodiment, each fully connected layer of the set of fully connected layers is configured to determine a correlation within the set of features of the three-dimensionally aligned facial image.

In one embodiment, the output of one fully connected layer of the set of fully connected layers is a feature vector.
In one embodiment, the DNN is configured to classify 2D face images based on feature vectors. The feature vector comprises a set of features of a 3D aligned face image. Each feature in the set of features is normalized to a predetermined range.

In one embodiment, the DNN is configured to define a set of filters for each pixel of the 3D aligned face image.
In one embodiment, the DNN is trained to define the set of filters using a data set.

  In one embodiment, the individual of the 2D face image may be identified by querying the ID of the 2D face image in the image store. The image store comprises a set of face images, and each face image in the set of face images corresponds to an individual.

  In one embodiment, the second ID of the second two-dimensional face image is classified. The ID can be compared with the second ID to determine whether the ID and the second ID correspond to an individual.

In one embodiment, a weighted χ ² distance between the ID and the second ID may be determined to compare the ID against the second ID.

  In one embodiment, the ID comprises a first set of features and the second ID comprises a second set of features. A set of feature differences between the first set of features and the second set of features may be determined. The set of feature differences can be mapped to a predetermined value by using the full coupling layer to compare the ID against the second ID.

  In a further embodiment according to the invention, one or more computer-readable non-transitory storage media, when executed, perform the method according to the invention or any of the above embodiments. Implement software that is operable.

  In a further embodiment according to the invention, the system comprises one or more processors and at least one memory coupled to the processor and comprising instructions executable by the processor, the processor in executing the instructions It is operable to perform any of the methods according to the invention or the above embodiments.

  In a further embodiment according to the present invention, preferably a computer program product comprising a computer-readable non-transitory storage medium is executed on a data processing system, the method according to the present invention or the above implementation. It is operable to perform any of the forms.

Many other features and embodiments of the present invention will become apparent from the following detailed description and from the accompanying drawings.
Embodiments according to the present invention are disclosed in the appended claims specifically directed to methods, storage media, systems and computer program products, and are described in one claim category, eg, method. Any feature may be claimed in another claim category, for example in a system. The dependencies or references in the appended claims are chosen solely for formal reasons. However, regardless of the dependencies selected in the appended claims, any preceding claim (especially in multiple dependent form) may be disclosed and claimed in any combination of claims and their features. Any subject matter derived from an intentional reference to a section can also be claimed. The claimed subject matter comprises not only the combination of features recited in the appended claims, but also any other combination of features in the claims, wherein each feature recited in the claims includes: It may be combined with any other feature in the claims, or with a combination of other features. Moreover, any of the embodiments and features described or illustrated herein may be claimed in separate claims and / or any described or illustrated herein. It may be claimed in any combination with the embodiments or features, or may be claimed in combination with any of the features of the appended claims.

Exemplary facial expressions configured to align facial images, classify facial images, and / or verify facial images by employing a deep neural network, according to an embodiment of the present disclosure FIG. 4 illustrates an example system that includes modules. FIG. 3 illustrates an exemplary face alignment module, according to one embodiment of the present disclosure. FIG. 4 illustrates an exemplary two-dimensional image comprising a detected face portion having a reference point, according to one embodiment of the present disclosure. FIG. 3B illustrates an example two-dimensional face image generated based on the example two-dimensional image shown in FIG. 3A, according to one embodiment of the present disclosure. FIG. 4 illustrates an exemplary two-dimensional face image having a detected reference point and a corresponding triangle defined by the reference point, according to one embodiment of the present disclosure. FIG. 4 illustrates an exemplary three-dimensional shape converted to an imaging plane of the two-dimensional face image shown in FIG. 3B, according to one embodiment of the present disclosure. FIG. 3 illustrates triangle visibility for a adapted 3D-2D camera according to an embodiment of the present disclosure. FIG. 4 illustrates anchor points arranged on a three-dimensional shape corresponding to the reference point illustrated in FIG. 3C, according to one embodiment of the present disclosure. FIG. 3 illustrates an exemplary three-dimensionally aligned face image according to one embodiment of the present disclosure. FIG. 3D is an exemplary view of a face generated based on the three-dimensional face image illustrated in FIG. 3G according to an embodiment of the present disclosure. FIG. 4 illustrates an exemplary face classification module, according to one embodiment of the present disclosure. FIG. 3 illustrates an example architecture of an example DNN, according to one embodiment of the present disclosure. FIG. 3 illustrates an exemplary method for generating a facial representation for a two-dimensional facial image according to an embodiment of the present disclosure. FIG. 3 illustrates an exemplary method for generating a facial representation for a two-dimensional facial image according to an embodiment of the present disclosure. FIG. 6 illustrates a receiver operating characteristic (“ROC”) curve for some embodiments compared to other methods for an LFW dataset. FIG. 6 illustrates a receiver operating characteristic (“ROC”) curve of some embodiments compared to other methods for a YTF data set. 1 is a network diagram of an example system that can be utilized in various scenarios according to one embodiment of the present disclosure. FIG. FIG. 3 illustrates an example computer system that can be utilized in various scenarios according to an embodiment of the present disclosure.

  The drawings are merely illustrative of various techniques of the disclosed technology for illustrative purposes, and the drawings use like reference numerals to identify like elements. Those skilled in the art will readily recognize from the following discussion that alternative embodiments of the structures and methods illustrated in the drawings may be employed without departing from the principles of the disclosed technology described herein. Will do.

  In recent years, many photos have been crawled by search engines and uploaded to social networks. These photographs may include a variety of unconstrained materials such as objects, faces, and scenes. The availability of this enormous amount of data is important for the computer vision community when dealing with its open issues. However, many of the generalization capabilities of traditional machine learning tools used in computer vision, such as support vector machines, principal component analysis, and linear discriminant analysis, have increased as the amount of training sets has increased significantly. Tend to saturate fairly quickly.

  Modern face recognition may comprise several stages such as detection, alignment, representation, and classification. During the detection stage, one or more faces may be detected from one image, and the detected faces are aligned in the alignment stage. Thereafter, each face in the image is represented during the presentation phase. Each face in the image can be classified in a classification stage. Conventional face recognition methods use manually selected features that are defined and selected by the user. Manually selected features can be combined to improve performance, even in the earliest Labeled Faces in the Wild ("LFW") contributions. Such a system may employ tens of thousands of image descriptors.

  Face alignment locates semantic face landmarks such as both eyes, nose, mouth, and chin. As a result, face alignment can be an important part of a face recognition system. A face database (eg, LFW-a) that includes an aligned face data set may improve the performance of the face recognition system by providing a normalized input. Some of the contributions on the LFW results page use a two-dimensionally aligned version of the dataset (eg, LFW-a) to improve the results. Other contributions may employ more customized solutions. For example, a personal data set of similar poses and expressions is used to infer a reasonable “ID preservation” alignment. Nevertheless, aligning faces in unconstrained scenarios can be difficult. Due to the non-planarity of the face, many factors (eg, posture and soft facial expressions) that hinder facial morphology and are difficult to separate from a person's ID need to be considered.

  FIG. 1 is an exemplary configuration configured to align face images, classify face images, and / or confirm face images by employing a deep neural network according to an embodiment of the present disclosure. 1 illustrates an exemplary system 100 that includes a simple facial expression module 102. As shown in FIG. 1, the exemplary face representation module 102 may include a face alignment module 104, a face classification module 106, and a face confirmation module 108.

  The example system 100 may include at least one image store 110. In the example of FIG. 1, at least one image store 110 may be configured to store images or other digital media such as video. The image may include an identified face image or an unidentified face image. The identified face image is linked to a person, while the unidentified face image is not linked to any person. For example, the identified face image may be tagged against a person ID (eg, name, user name, etc.). The image may be two-dimensional or three-dimensional. In various embodiments, the original image can be used as the underlying representation. The image store 110 may store a data set that may be used to train a deep neural network (“DNN”) to learn and classify facial expressions. Various embodiments provide a deep neural network architecture and learning method that leverages very large labeled datasets of faces to obtain facial expressions that are sufficiently generalized to other datasets. Can be provided. A DNN may be involved, for example, in over 120 million parameters using several local coupling layers without weight sharing. For example, the data set may comprise 4 million face images belonging to more than 4000 IDs. Here, each ID has a sample exceeding an average of one person.

In some embodiments, the face alignment module 104 may be configured to include analytical three-dimensional modeling of the face based on reference points. The face alignment module 104 is configured to detect a detected face crop (facial) based on an analytical 3D model.
crop) and thereby can be further configured to create a three-dimensionally aligned face image. Since the position of each face region is fixed at the pixel level in the three-dimensionally aligned face image, the facial appearance can be learned and detected from the RGB values of the pixels in the original image.

  In some embodiments, the face classification module 106 may be configured to classify image IDs. Face classification module 106 may be configured to classify an image ID (eg, facial expression) based directly on the RGB values of the pixels of the image. Image IDs can be very short and even sparse. In further embodiments, the image ID may be generated based on the RGB values of the pixels, the gradient domain of the image, the depth map, or depth information, or any combination thereof.

In some embodiments, the face verification module 108 is configured to verify whether two images have the same ID. In other words, it is confirmed whether two images belong to the same class. Face verification module 108 may be configured to determine the similarity between two images based on the ID (eg, facial representation of the image) classified by face classification module 106. An image can be represented by a feature vector. In some embodiments, the similarity of two images can be determined by calculating an inner product between two normalized feature vectors for the two images. In some embodiments, the similarity between two images may be determined by a monitored metric such as χ ² similarity or a Siamese network.

The weighted χ ² similarity between two images is χ ² (f ₁ , f ₂ ) = Σ _i w _i (f ₁ (i) −f ₂ (i)) ² / (f ₁ (i) + f ₂ (i)). Here, f ₁ and f ₂ are each a feature vector representation of the image. The weight parameter can be learned using linear SVM applied to a vector of elements (f ₁ (i) −f ₂ (i)) ² / (f ₁ (i) + f ₂ (i)). .

The Siamese network provides a highly nonlinear metric for two images. The absolute difference between the features identified for the two images can be determined and used to directly predict whether the two input images belong to the same person. The induced distance between features for two images can be expressed as d (f ₁ , f ₂ ) = Σ _i a _i | f ₁ (i) −f ₂ (i) |. Here, a _i is a trainable parameter of the Siamese network. The parameters of the Siamese network ( _ai and joint parameters in the lower layers) are determined by normalizing the distance between 0 and 1 through the logistic function 1 / (1 + exp (−d)), as well as cross entropy It can be trained by using loss and back propagation. In various embodiments, training is only possible for the two top layers of the Siamese network to prevent overfitting for face verification tasks.

  FIG. 2 illustrates an exemplary face alignment module 202 as shown in FIG. 1 (eg, face alignment module 104), according to an embodiment of the present disclosure. The example face alignment module 202 may align face images. Images to be aligned may be obtained from at least one data store 208 (eg, image store 110 in FIG. 1). The aligned face image can be stored in the data store 208. FIG. 2 is discussed in connection with FIGS. 3A-3H.

  3A-3H illustrate exemplary image data at different times of the alignment process, according to embodiments of the present disclosure. FIG. 3A illustrates an exemplary two-dimensional image 301 comprising a detected face portion 310 having a reference point. FIG. 3B illustrates an exemplary two-dimensional face image 302 generated based on the exemplary two-dimensional image 301 illustrated in FIG. 3A. FIG. 3C illustrates an exemplary two-dimensional face image 303 having a detected reference point and a corresponding triangle defined by the reference point. FIG. 3D illustrates an exemplary three-dimensional shape 304 that has been converted to the imaging plane of the two-dimensional face image 302 illustrated in FIG. 3B. FIG. 3E illustrates triangle visibility for an adapted 3D-2D camera. FIG. 3F illustrates anchor points placed on the three-dimensional shape 306 corresponding to the reference points illustrated in FIG. 3C. FIG. 3G illustrates an exemplary three-dimensional aligned face image 307. FIG. 3H illustrates an exemplary view of a face 308 generated based on the three-dimensional face image 307 illustrated in FIG. 3G.

  The face alignment module 202 may include a reference point detection module 204 and an image conversion module 206. The face alignment module may be configured to align the 2D image 210 to the 3D aligned face image 212. The three-dimensionally aligned face image 212 can be aligned in an unconstrained scenario so that many factors that hinder facial morphology and are difficult to separate from the person's ID can be considered. For example, the posture due to the non-planarity of the face and the soft facial expression is considered.

The reference point detection module 204 localizes the reference points in the two-dimensional face image. The two-dimensional face image can be an unconstrained image. In various embodiments, the reference points may be extracted by a support vector regressor (SVR). The SVR can be trained to predict the point configuration from the image descriptor of the 2D face image. In various embodiments, the image descriptor is based on a local binary pattern (“LBP (Local Binary Patterns)”) histogram. In further embodiments, other features may be considered in the image descriptor. In some embodiments, multiple iterations can be applied so that the reference point detection module 204 can improve its output. That is, the reference point detection module 204 can be configured to detect an initial set of reference points. The initial set of reference points can be refined to a final set of reference points through multiple iterations. The two-dimensional face image can be converted into a two-dimensional converted image using the derived similarity matrix T. The reference point detection module 204 can analyze the two-dimensional transformed image on a new feature space that can be used to improve the localization of the reference points of the two-dimensional face image.

The reference point detection module 204 can be configured to detect a set of reference points of the two-dimensional face image. Referring to FIG. 3C, an exemplary two-dimensional face image 303 having a set of reference points is illustrated. The set of reference points x _2d is detected from the two-dimensional face image 302 illustrated in FIG. 3B, which is a two-dimensionally aligned crop image of the two-dimensional image 301 illustrated in FIG. 3A. A set of reference points x _2d can be detected by using a trained SVR. In the illustrated example, a selected number (eg, 67) reference points are localized on the two-dimensional face image 303. The set of reference points x _2d may define a corresponding set of triangles (eg, Delaunay triangulation). A corresponding set of triangles can be added to the contour of the image illustrated in the two-dimensional face image 303. Therefore, cropping and acute angles can be avoided.

  Referring back to FIG. 2, the image conversion module 206 is three-dimensionally aligned by processing the two-dimensional face image into a three-dimensional shape by using the set of reference points detected by the reference point detection module 204. A face image 212 may be configured to be generated. Referring to FIG. 3G, an example three-dimensionally aligned face image 307 created based on the two-dimensional face image 302 illustrated in FIG. 3B is illustrated. The 2D face image 302 may be processed using the set of reference points illustrated in FIG. 3C to generate the 3D aligned face image 307 illustrated in FIG. 3G. When processing a two-dimensional face image into a three-dimensional shape or a three-dimensional model, the position of each reference point from the two-dimensional face image is back-projected to the three-dimensional shape or the three-dimensional model through the restored projection matrix P, and then It is projected onto the front image of the 3D shape or 3D model. In various embodiments, a piecewise affine transformation is used that is directed by a triangle (eg, Delaunay triangulation) derived from a set of reference points detected on a two-dimensional face image.

Referring to FIG. 3D, an exemplary three-dimensional shape or three-dimensional model 304 converted to the image plane of the two-dimensional face image 302 is illustrated. A general 3D shape or 3D model can be used as a reference for aligning the face image 302 including external plane rotation. In various embodiments, a typical 3D shape model is an average of 3D scans from the USF Human ID database. The average value of the 3D scan can be processed and represented as aligned vertices v _i = (x _i , y _i , z _i ) ⁿ _i = 1.

The set of anchor points X _3d can be arranged on a three-dimensional shape. Each anchor point corresponds to a reference point detected on the two-dimensional face image. A perfect match between a reference point detected on the 2D face image and an anchor point on the 3D aligned face image can be achieved. For example, FIG. 3F illustrates anchor points that can be derived by a three-dimensional shape or a three-dimensional model. Each of the anchor points illustrated in FIG. 3F corresponds to a reference point illustrated in FIG. 3C. The reference points and anchor points can be used to direct the piecewise affine processing of the 2D face image illustrated in FIG. 3B into the 3D aligned face image illustrated in FIG. 3G.

The 3D affine camera P is registered and used to backproject the front face plane of the 2D face image 302 illustrated in FIG. 3B onto the image plane of the 3D shape 304 illustrated in FIG. 3D. obtain. Therefore, the reference three-dimensional shape 304 is converted into the image plane of the two-dimensional face image 302. In other words, the two-dimensional face image 302 is converted into a three-dimensionally aligned face image based on the reference three-dimensional shape 304. Affine 3D-2D camera P

Can be solved using the least squares method with known covariances. Here, X _3d is the reference reference point position

(2 × 8) matrix stack given by

Represents four zero row vectors. The size 2x4 affine camera P has 8 unknown

Can be represented by a vector of In some embodiments, a (67 * 2) × (67 * 2) covariance matrix Σ given by the estimated covariance of the reference point error is used. This is because detected points on the face contour can be more noisy because their estimated position can be affected by depth relative to the camera angle.

  FIG. 3E illustrates triangle visibility for a fitted affine 3D-2D camera. The visibility of a triangle on a 3D shape or 3D model may depend on the distance of the corresponding anchor point that defines the triangle from the affine 3D-2D camera. Depending on the distance of the reference point from the camera, the triangle 334 is less visible than the triangle 332, and the triangle 332 is less visible than the triangle 330.

Referring back to FIG. 2, in various embodiments, the face alignment module 202 can be configured to generate a 2D face image from the 2D image 210. Referring to FIG. 3A, the face portion 310 can be detected from the two-dimensional image 301. The face portion 310 of the two-dimensional image 301 can be identified by detecting a set of reference points 312, 314, 316, 318, 320, and 322. Reference points 312, 314 may be centered on the center of the eye, reference point 316 may be centered on the tip of the nose, and reference points 318, 322, and 320 may be centered on both ends and center of the mouth, respectively. The reference points detected in the face portion can be used to determine a set of anchor positions. The identified face portion can be processed based on a set of anchor positions to generate a two-dimensional face image. For example, as illustrated, the reference points 312, 314, 316, 318, 320, and 322 may be used to scale the face portion 310 approximately to three anchor positions, rotate, and convert. To generate a new processed image as illustrated in FIG. 3B, the three anchor positions can be determined by solving for T ⁱ _2d : = (s _i , R _i , t _i ). Here, the point j = 1. . For n, x ^j _anchor : = s _i [R _i | t _i ] * x ^j _source . FIG. 3B illustrates a two-dimensional face image 302. The 2D face image 302 may be an aligned face crop generated by an aggregated transformation based on the 2D image 301. Multiple iterations are performed to determine the anchor position until there is no substantial change, thereby generating a two-dimensional face image 302 based on the detected face portion 310. Final two-dimensional similarity transformation T _2d : = T ^J _2d *. . . * T ^K _2d may be configured.

  FIG. 4 illustrates an exemplary face classification module 402, as shown in FIG. 1 (eg, face classification module 106), according to an embodiment of the present disclosure. The exemplary face classification module 402 may classify the ID of the image 412. Images to be classified may be obtained from at least one data store 410 (eg, image store 110 in FIG. 1). The classified images can be stored in the data store 410. FIG. 4 is discussed in connection with FIG.

  The face classification module 402 may include a filter generation module 404, a feature vector generation module 406, and an expression categorization module 408. The face classification module 402 may be configured to obtain an image ID (eg, facial expression) from the DNN. The filter generation module 404 may be configured to generate a set of filters by training the DNN with the data set. In various embodiments, the data set may include a large number of photos. A portion of the photo may be identified (eg, labeled or tagged with an ID). Each filter may be configured to extract image features (eg, facial features). The feature vector generation module 406 may be configured to generate a feature vector for the image. The feature vector includes various features of the image and can be used to represent the image. An image feature vector may be generated based on the set of filters provided by the filter generation module 404. The representation categorization module 408 may categorize the images based on the image feature vectors.

  In various embodiments, the DNN comprises a set of layers, the set of layers including a convolutional pooling layer, a local tie layer, and a full tie layer. The convolution pooling layer may be configured to extract a set of image features. The local coupling layer may be configured to extract another set of image features. The fully connected layer may be configured to determine a correlation between image features. In some embodiments, the overall architecture of the DNN includes a convolutional pooling layer followed by a set of local coupling layers, followed by a full coupling layer. The DNN may comprise at least a selected number (eg, 120 million) of parameters (eg, connection weights). In various embodiments, the DNN is configured to receive as input a 3D aligned face image, such as the 3D aligned face image 307 illustrated in FIG. 3G.

In various embodiments, DNN uses the ReLU activation function max (0, x). This soft-thresholding non-linearity can be applied after the convolutional layer, the local coupling layer, or the total coupling layer, except for the last total coupling layer of the DNN. Thus, highly nonlinear and sparse features can be created. Thus, the features produced by the face classification module 402 can be very sparse. In some embodiments, on average, 75% of the feature components in the top layer are just zero. In further embodiments, a regularization method (eg, dropout) that sets random feature components to 0 during training may be applied. In some embodiments, the regularization method may be applied only to the first fully connected layer. The use of ReLU activation allows the face classification module 402 to remain invariant to image intensity adjustment. Without bias in DNN, complete equivalence, or almost complete homogeneity can be achieved.

In some embodiments, the face classification module 402 may be configured to normalize features to be between zero and one. Face classification module 402 may be further configured to perform the L ₂ normalized. Therefore, subsequent learning can be improved. Each component of the feature vector can be divided by its maximum value over the entire training set. In some embodiments, the normalization factor is capped at a predetermined value (eg, 0.05) to avoid division by a small number. If the maximum number is less than a predetermined value (eg, 0.05), the features are not normalized. Such normalization may ensure that the face classification module 402 is robust to lighting changes.

  In various embodiments, the input 3D aligned face image may be provided to the DNN convolution pooling layer. The convolution pooling layer may comprise a first convolution layer (C1), a maximum pooling layer (M2), and a second convolution layer (C3). Each convolution layer may be configured with a set of filters. Each of the filters is configured to generate a feature map. The maximum pooling layer can make DNN more robust to small registration errors. However, some levels of pooling can cause the network to lose information about the detailed facial structure and the exact location of the micro-texture. Thus, in various embodiments, maximum pooling can be applied only to the first convolution layer. That is, the maximum pooling layer follows only after the first convolution layer. The convolution pooling layer is configured to extract low-level features of the image (eg, simple edges and textures). The convolution pooling layer adaptively preprocesses the input image and extends it to a simple set of local features.

  FIG. 5 illustrates an example architecture 500 of an example DNN, according to an embodiment of the present disclosure. The three-dimensionally aligned face image 524 (including RGP copies 524a-524c) is provided to DNN 500 as input. In the illustrated example, the DNN 500 may classify the ID of the three-dimensionally aligned face image 514 based on the RGB values of the pixels. Local facial features of image 524 may be extracted by convolution pooling layer 502 and local coupling layer set 510. The correlation between local facial features of the image 524 may be determined by the set of all connected layers 518. In the illustrated example, the output of all connected layers F7 520 may be used as the feature vector of image 524.

As illustrated, the exemplary convolution pooling layer 502 comprises a convolution layer C1 504, a maximum pooling layer M2 506, and a convolution layer C3 508. The convolution layer may be configured to have a set of filters (eg, 32 filters). In one embodiment, each filter of the selected number of filters may be 11x11x3 in size. The set of feature maps generated by the convolution layer C1 can be provided to the maximum pooling layer M2. For each channel, the maximum pooling layer M2 may be configured to take a maximum value at stride 2 for a 3 × 3 space neighborhood. The second convolution layer C3 can be applied to the output of the maximum pooling layer M2. The convolutional layer C3 may be configured to have a set of filters (eg, 16 filters). In one embodiment,
Each filter of the selected number of filters may be 9x9x16 in size.

  Image features (eg, local facial features) generated by the convolutional pooling layer may be provided to a set of local coupling layers. The set of local coupling layers does not share weights, and each output of the local coupling layer can be affected by very large patches of input. Thus, the use of a local coupling layer may not affect the computational load of feature extraction, but may affect the number of parameters that are trained. Each local coupling layer may be configured to extract a set of features from the input image. Different features can be extracted by the local combined layer and the convolutional layer for the same location of the image. That is, the set of filters at every location in the feature map is learned by the DNN local coupling layer, which is different from the set of filters learned by the convolutional layer. The local coupling layer does not make the assumption of spatial stationarity of convolution. This is because different regions of the aligned image can have different local statistics. For example, the area between both eyes and both eyebrows presents a very different appearance and has a much higher discriminating power compared to the area between the nose and mouth.

  Referring also to FIG. 5, as illustrated, an exemplary set 510 of local tie layers includes a first local tie layer L4 512, a second local tie layer L5 514, and a third A local coupling layer L6 516. At every position in the input image feature map for the exemplary DNN 500 illustrated, the local coupling layer L4 512, L5 514, or L6 516 is a filter from the set applied by the convolutional layer C1 504 or C3 508. Different sets of can be applied. Each output unit of the local coupling layer L4 512, L5 514, or L6 516 may be affected by a very large patch of input. For example, as illustrated, the output of the local coupling layer L6 516 may be influenced by a 74x74x3 patch of the input image 524. Since the input image 524 is aligned, there is almost no statistical sharing between such patches.

  The local facial features of the image generated by the convolution pooling layer and the set of local coupling layers can be provided to the set of all coupling layers. The set of fully connected layers may be configured to determine a correlation between local facial features of the image generated by the convolutional pooling layer and the set of locally connected layers. That is, the set of all connected layers is configured to determine the correlation between local facial features (eg, eye position and shape, mouth position and shape) extracted in distant parts of the face image. The Each output of the total coupling layer can be connected to all inputs of the total coupling layer.

Referring to FIG. 4, in some embodiments, feature vector generation module 406 may be configured to use the output of all connected layers as the feature vector of image 410. The expression categorization module 408 may be configured to categorize images into classes based on feature vectors. In some embodiments, the output of the DNN (e.g., the output of the last full coupling layer) may be provided to a K-way softmax. Here, K is the number of classes. The K-direction softmax can be configured to produce a variance across class labels. The probability assigned to the i th class is the output of the softmax function p _i = exp (o _i ) / Σ _j exp (o _j ), where o _i is the network for a given output Represents the i-th output of. The expression categorization module 408 can be configured to identify the individual to which the two-dimensional face image belongs. The image store 410 can be queried using the ID (eg, facial expression, feature vector) of the two-dimensional face image.

In various embodiments, when training a DNN, the probability of the correct class (eg, face ID) is maximized. The cross entropy loss for each training sample can be minimized to maximize the probability of the correct class. The cross entropy loss for a given input is L = −logp _k , where k is the true label index for the given input. By calculating the slope of L for the parameter and by updating the parameter using a stochastic gradient descendant (SGD), the loss can be minimized across the parameter. In some embodiments, the slope may be determined by standard back propagation of errors.

  The architecture of DNN 500 can be customized by taking advantage of the fact that the input images are aligned. The exemplary DNN 500 supports various facial features to detect various facial features and / or generate image feature vectors (as described in more detail above). Can be configured to learn different filters. As illustrated in FIG. 5, the exemplary DNN 500 is a nine-layer DNN comprising a convolutional pooling layer 502, a set of local coupling layers 510, and a set of all coupling layers 518. The convolution pooling layer 502 may comprise a convolution layer C1 504, a maximum pooling layer M2 506, and a convolution layer C3 508. The set of local tie layers 510 can comprise local tie layers L4-L6 512, 514, and 516, and the set of all tie layers 518 can comprise all tie layers F7-F8 520 and 522. Each output unit of all coupling layers F7 520 or F8 522 may be connected to all inputs.

  FIG. 6 illustrates an exemplary method for generating a facial representation for a two-dimensional facial image according to an embodiment of the present disclosure. Unless otherwise stated, within the scope of the various embodiments, there are additional steps, fewer steps, or alternative steps that are performed in a similar or alternative order, or performed in parallel. It should be recognized that it can exist. At block 602, a three-dimensionally aligned face image is generated based on the two-dimensional face image. A 3D aligned face image can be created by applying an affine transformation of a 2D face image based on a 3D model. The 3D model can be a clear 3D face modeling. In block 604, the ID of the 2D face image may be classified by providing the 3D aligned face image to the DNN. A facial representation (eg, a feature vector) of a two-dimensional facial image can be generated by using a three-dimensionally aligned facial image that corresponds to the two-dimensional facial image. In some embodiments, the DNN may comprise nine layers. A DNN may be involved, for example, in over 120 million parameters using multiple local coupling layers without weight sharing. More detailed discussion and examples are provided above with reference to FIGS.

  FIG. 7 illustrates an exemplary method for generating a facial representation for a two-dimensional facial image, according to an embodiment of the present disclosure. Unless otherwise stated, within the scope of the various embodiments, there are additional steps, fewer steps, or alternative steps that are performed in a similar or alternative order, or performed in parallel. It should be recognized that it can exist. At block 702, a face portion in the two-dimensional image is identified by detecting a set of reference points on the two-dimensional image. The reference point can be detected by an SVR trained to predict the point configuration from the image descriptor. In block 704, a two-dimensional face image may be generated by transforming the face portion of the two-dimensional image based on the reference points detected in block 702. A two-dimensional face image can be an aligned crop of a two-dimensional image and can be generated by an aggregated transform. When the 2D similarity transformation is finally configured, the reference points are scaled, rotated, transformed, and reprocessed until no substantial change exists, the face portion of the 2D image Can be used to iterate over the image.

At block 706, a different set of reference points may be detected in the two-dimensional face image. The reference point may be localized by using a SVR that is different from the SVR used in block 702. In block 708, the two-dimensional face image may be processed into a three-dimensional shape by using the reference points detected in block 706 to generate a three-dimensionally aligned face image. The three-dimensional shape can be a general three-dimensional model. A three-dimensional affine camera may be registered, and the three-dimensional affine camera is registered and used to backproject the front face plane of the two-dimensional face image onto a three-dimensional shaped imaging plane, thereby A dimensionally aligned face image is generated. At block 710, the ID of the 2D face image may be classified by providing the 3D aligned face image to a deep neural network (“DNN”). More detailed discussion and examples are provided above with reference to FIGS.

  In various embodiments, the DNN is configured to learn facial expressions from a large number of photos, referred to as social face classification ("SFC") data sets. The representation may be applied to the LFW database and Youtube (R) face ("YTF"). The LFW database can be used as a de facto benchmark data set for face recognition in an unconstrained environment. YTF datasets can be modeled in the same way as LFW, but focus on video clips. Face IDs in the SFC are labeled by humans and typically contain about 3% errors. Social face photos have greater variations in image quality, lighting, and facial expression than celebrity web images in LFW and YTF, which are usually taken by professional photographers rather than smartphones.

  The SFC data set as used here contains 4.4 million labeled faces from 4,030 people. Each individual may have 800 to 1200 facial images. The latest 5% of the face image of each ID can be omitted. The LFW data set as used here consists of 13,323 web photos of 5,749 celebrities, which are divided into 6,000 face pairs divided into 10 parts. The performance uses a limited protocol where only the “same” and “not the same” labels are available in training, or an unrestricted protocol where the training target ID is also accessible in training. Measured by average recognition accuracy. Also, the “no monitoring” protocol may measure performance on the LFW without training for this data set.

  A YTF data set as used here collects 3,595 Youtube® videos of 1,595 themes, which may be a subset of celebrities in the LFW. These videos are divided into 5,000 video pairs, divided into 10 and can be used to evaluate video level face recognition.

  DNN is trained on SFC by implementing standard error backpropagation for feed forward nets by stochastic gradient descent (SGD) with momentum (set to 0.9). obtain. The mini batch size is 128, and an equal learning rate of 0.01 is ensured for all learning layers. The learning rate can be manually reduced to a final rate of 0.00001. Once confirmation errors no longer decrease, the learning rate can be reduced by an order of magnitude each time. The weights in each layer can be initialized from a mean zero Gaussian distribution with o = 0.01 and bias set to 0.5. The network can be trained for approximately 15 sweeps (epochs) throughout the data.

Table 1 illustrates the classification error of an embodiment on the SFC with respect to training dataset size and network depth. Training and test dataset sizes were changed using a subset of individuals in SFC. For SFCs with 1.5 million, 3.3 million, and 4.4 million faces, a subset of the size of 1500, 3000, and 4000 individuals, respectively, was used. Multiple DNNs were trained (e.g., DeepFace-1.5 million, DeepFace-3.3 million, and DeepFace-4.4 million). As illustrated, large face data sets and deep architectures reduce error rates. The classification error only rises moderately from 7.0% for 1500 people to 7.2% when 3,000 people are classified. This indicates that the capacity of the network can sufficiently cope with the scale of 3 million training images. The error rate has risen to 8.7% for 4000 people with 4.4 million images, indicating that the network scales reasonably to more individuals.

The depth of the network is changed by removing one or more layers from the architecture illustrated in FIG. For example, the convolution layer (eg, C3) is removed from DNN DeepFace-sub1, the set of local coupling layers (eg, L4 and L5) is removed from DNN DeepFace-sub2, and the convolutional layer (eg, C3) and multiple The local tie layer (eg, L4 and L5) is removed from DNN DeepFace-sub3. DNN is trained with 4.4 million faces. Classification errors no longer decrease after several epochs and remain at a level higher than that of the deep network DeepFace-44.4 million.

FIG. 8A and Table 2 illustrate the receiver operating characteristic (“ROC”) curve and confirmation accuracy of some embodiments compared to other methods for the LFW data set. Curves 802, 804, 806, 808, 810, 812, and 814 are respectively in an embodiment using human confirmation, ensemble confirmation (eg, DeepFace-ensemble), in a 3D aligned face on an SFC data set Embodiments with DNNs learned from raw RGB pixels (eg, Deepface-single), TL joint Bayesian, high dimensional LBP, Tom-vs-Pete, and double coupled Bayesian (combined) joint
Bayesian). DeepFace-ensemble is trained by supplying different types of input to DNN, including 3D aligned RGB input, halftone image and image gradient magnitude and orientation, and 2D aligned RGB image. Combine multiple network distances. The distances are combined using a non-linear SVM (C = 1) with a simple sum of exponents CDF-kernels: K _Combined : = K _single + K _gradient + K _{align 2d} . Here, K (x, y): = − || xy− | ₂ .

The unsupervised protocol is observed to directly compare the dot product of normalized feature pairs. The restricted protocol uses 5,400 pairs of labels per partition for SVM training. The unrestricted protocol allows the creation of more training pairs to be added to the training set to provide the operator with knowledge about the IDs in the training set. As illustrated in FIG. 8A and Table 2, the embodiments described herein advance other methods and closely approach human performance in face recognition.

  FIG. 8B and Table 3 illustrate the receiver operating characteristic (“ROC”) curve and confirmation accuracy of some embodiments compared to other methods for a YTF data set. Curves 852, 854, 856, 858, 860, 862 are embodiments with DNNs learned from raw RGB pixels in a three-dimensionally aligned face against the SFC dataset, respectively (eg DeepFace-single), VSOF + OSS, STFRD + PMML. , APEM + FUSION, MBGS (average) LBP, and MBGS (average) FPLBP.

The image quality of YouTube® video frames is generally lower than that of web photos, mainly due to subject blurring or viewing distance. For every pair of training videos, one is taken from each video to create fifty (50) pairs of frames, which are the same or not the same according to the video training pair Labeled as A weighted χ ² model is learned. Given a test pair, one hundred (100) random pairs of frames are sampled by sampling one from each video and use the learned weighted similarity average.

  As illustrated, the embodiments described herein provide 91.4% accuracy and reduce past best method errors by over 50%. The accuracy can be improved to 92.5% after correcting 100 incorrect labels for video pairs in the YTF data set.

Social Networking System-Exemplary Implementation FIG. 9 illustrates a network diagram of an exemplary system 900 that may be utilized in various embodiments for enhanced video encoding, according to embodiments of the present disclosure. System 900 includes one or more user devices 910, one or more external systems 920, a social networking system 930, and a network 950. In one embodiment, the social networking services, providers, and / or systems discussed in connection with the above-described embodiments may be implemented as a social networking system 930. For illustrative purposes, the embodiment of system 900 illustrated by FIG. 9 includes a single external system 920 and a single user device 910. However, in other embodiments, the system 900 may include more user devices 910 and / or more external systems 920. In certain embodiments, the social networking system 930 is operated by a social network provider, whereas the external system 920 can be operated by different entities, Are separate. However, in various embodiments, the social networking system 930 and the external system 920 work together to provide social networking services to users (or members) of the social networking system 930. In this sense, social networking system 930 provides a platform or backbone that other systems, such as external system 920, can use to provide social networking services and functionality to users on the Internet. .

  User device 910 comprises one or more computing devices that can receive input from a user and send and receive data over network 950. In one embodiment, the user device 910 is, for example, a conventional operating system (OS), Apple OS X, and / or Linux distribution that is compatible with Microsoft Windows. Computer system. In another embodiment, the user device 910 may be a device having computer functionality, such as a smartphone, a tablet, a personal digital assistant (PDA), a mobile phone, and the like. User device 910 is configured to communicate over network 950. User device 910 may execute an application, eg, a browser application that allows a user of user device 910 to interact with social networking system 930. In another embodiment, the user device 910 is a social networking system through an application programming interface (API) provided by the native operating system of the user device 910, such as iOS and ANDROID®. 930 interacts. User device 910 is configured to communicate with external system 920 and social networking system 930 over network 950. Network 950 may comprise any combination of local area networks and / or wide area networks using wired and / or wireless communication systems.

In one embodiment, the network 950 uses standard communication technologies and protocols. Accordingly, the network 950 includes Ethernet (registered trademark) 702.
11. Global interoperability for microwave access (WiMAX: worldwide)
It may include links using technologies such as interoperability for microwave access), 3G, 4G, CDMA, GSM®, LTE, digital subscriber line (DSL), and the like. Similarly, networking protocols used on the network 950 include multiprotocol label switching (MPLS), transmission control protocol / Internet protocol (TCP / IP), user datagram protocol (UDP). ), Hypertext transfer protocol (HTTP), simple mail transfer protocol (SMTP), file transfer protocol (FTP), and the like. Data exchanged over the network 950 may be represented using techniques and / or formats including hypertext markup language (HTML) and extensible markup language (XML). In addition, all or some of the links are encrypted using conventional encryption techniques such as Secure Sockets Layer (SSL), Transport Layer Security (TLS), and Internet Protocol Security (IPsec). Can be

  In one embodiment, the user device 910 processes the markup language document 914 received from the external system 920 and from the social networking system 930 using the browser application 912 to provide an external system. Content from 920 and / or content from social networking system 930 may be displayed. Markup language document 914 identifies content and one or more instructions that describe how to format or display the content. By executing the instructions contained within the markup language document 914, the browser application 912 displays the identified content using the format or display method described by the markup language document 914. For example, markup language document 914 includes instructions for generating and displaying a web page having a plurality of frames containing text and / or image data obtained from external system 920 and social networking system 930. In various embodiments, the markup language document 914 comprises a data file that includes extended markup language (XML) data, extended hypertext markup language (XHTML) data, or other markup language data. In addition, the markup language document 914 includes JavaScript object notification (JSON) data and padding for facilitating data exchange between the external system 920 and the user device 910. JSON (JSONP: JSON with padding) data, and Javascript data. Browser application 912 on user device 910 may decrypt markup language document 914 using a JavaScript compiler.

  The markup language document 914 also includes or may be linked to an application or application framework, such as a FLASH ™ application or a Unity ™ application, a SilverLight ™ application framework, etc. .

  In one embodiment, user device 910 also includes one or more cookies 916 that include data indicating whether the user of user device 910 has logged into social networking system 930. Cookie 916 may allow modification of data communicated from social networking system 930 to user device 910.

  External system 920 includes one or more web servers that include one or more web pages 922a, 922b. One or more web pages 922 a, 922 b are communicated to user device 910 using network 950. External system 920 is separate from social networking system 930. For example, the external system 920 is associated with a first domain while the social networking system 930 is associated with a separate social networking domain. Web pages 922a, 922b included in external system 920 include a markup language document 914 that includes instructions for identifying content and specifying how to format and display the identified content.

  Social networking system 930 includes a plurality of users and provides a social network user with the ability to communicate and interact with other users of the social network. Including one or more computing devices. In some examples, a social network may be represented by a graph, ie, a data structure that includes edges and nodes. Other data structures may also be used to represent social networks, including but not limited to databases, objects, classes, meta elements, files, or any other data structure. Social networking system 930 may be operated, managed, or controlled by an operator. The operator of the social networking system 930 manages the content within the social networking system 930, regulates policies, collects usage metrics, humans, automated applications, or a set of It may be an application. Any type of operator can be used.

  A user may subscribe to a social networking system 930 and then add a connection to any number of other users of the social networking system 930 that the user wishes to be connected to. As used herein, the term “friend” refers to any other user of the social networking system 930 that the user has connected, formed an association or relationship through the social networking system 930. For example, in one embodiment, if a user in social networking system 930 is represented as a node in the social graph, the term “friend” is formed between two user nodes, and 2 It may refer to an edge that directly connects two user nodes.

  Connections may be added explicitly by the user or automatically created by the social networking system 930 based on the user's common characteristics (eg, users who are graduates of the same institution) Good. For example, the first user specifically selects a specific other user who wants to be a friend. The connection in social networking system 930 is usually bi-directional, but need not be bi-directional, so the terms “user” and “friend” depend on the framework. The connection between users of social networking system 930 is typically bilateral (“two-way”) or “reciprocal”, but the connection may be one-way or “one-way”. . For example, if both Bob and Joe are users of social networking system 930 and are connected to each other, Bob and Joe are connected to each other. On the other hand, if Bob wants to connect with Joe to see the data communicated by Joe to social networking system 930, but Joe does not want to form a reciprocal connection, Connections can be established. The connection between users can be a direct connection. However, some embodiments of the social networking system 930 allow connections to be indirect through one or more levels or degrees of connection.

  In addition to establishing and maintaining connections between users and enabling interaction between users, the social networking system 930 provides for various types of items supported by the social networking system 930. Provide users with the ability to take action. Such items include groups or networks to which users of social networking system 930 may belong (ie, social networks of people, entities, and concepts), events or calendar entries that may be of interest to users, and social networking of users. Computer-based applications that can be used through system 930, transactions that allow a user to purchase or sell items through services provided by social networking system 930 or through social networking system 930, and It may include interaction with advertisements that may run inside and outside the social networking system 930. These are just a few examples of items that a user can affect on the social networking system 930, and many other examples are possible. The user is represented in social networking system 930 or in external system 920 that is coupled to social networking system 930 through external system 920 or network 950 that is separate from social networking system 930. Can interact with everything possible.

  The social networking system 930 can also be linked with various entities. For example, social networking system 930 allows users to interact with each other and with external systems 920 or other entities through APIs, web services, or other communication channels. Social networking system 930 generates and maintains a “social graph” comprising a plurality of nodes interconnected by a plurality of edges. Each node in the social graph may represent an entity that can act on another node and / or an entity that can be acted on by another node. Social graphs can include various types of nodes. Examples of node types include users, non-human entities, content items, web pages, groups, activities, messages, concepts, and anything else that can be represented by objects in the social networking system 930. An edge between two nodes in a social graph can show a particular type of connection or association between two nodes that can result from a node relationship or from an action performed on one by one of the nodes. It can be expressed. In some cases, the edges between nodes may be weighted. An edge weight may represent an attribute associated with that edge, such as the strength of a connection or association between nodes. Different types of edges can be provided with different weights. For example, an edge created when a user asserts “Like” to another user may be given a weight while being created when the user makes a friend with another user The edges can be given different weights.

  As an example, if a first user identifies a second user as a friend, a node in the social graph connecting a node representing the first user and a second node representing the second user An edge is generated. As the various nodes relate to or interact with each other, the social networking system 930 changes the edges connecting the various nodes to reflect the relationships and interactions.

Social networking system 930 also includes user generated content that enhances user interaction with social networking system 930. Content generated by the user may include anything that the user can add to, upload, send, or “post” to the social networking system 930. For example, the user communicates a post from the user device 910 to the social networking system 930. Posts may include data such as status updates or other textual information, location information, images such as photos, videos, links, music, or other similar data and / or media. Content can also be added to the social networking system 930 by a third party. The content “item” is represented as an object in the social networking system 930. In this way, users of social networking system 930 are encouraged to communicate with each other by posting various types of text and content items over various communication channels. Such communication increases user interaction and increases the frequency with which the user interacts with social networking system 930.

  Social networking system 930 includes web server 932, API request server 934, user profile store 936, connection store 938, behavior logger 940, activity log 942, and authentication server 944. . In one embodiment of the present invention, social networking system 930 may include additional components, fewer components, or different components for various applications. Other components such as network interfaces, security mechanisms, load balancers, failover servers, management and network operations consoles are not shown so as not to obscure the details of the system.

  The user profile store 936 is a career history, educational background, hobbies or preferences, historical information such as address, demographic information, and other types of descriptions as declared by the user or inferred by the social networking system 930. Maintain information about user accounts, including tactical information. This information is stored in the user profile store 936 so that each user is uniquely identified. Social networking system 930 also stores data in connection store 938 describing one or more connections between different users. The connection information may indicate users with similar or common work history, group membership, hobbies, or educational background. The social networking system 930 also includes user-defined connections between different users that allow the user to specify the relationship of the user to other users. For example, a user-defined connection allows a user to create a relationship with another user, such as a friend, colleague, partner, etc., in parallel with the user's real-life relationship. The user may select from a predefined type of connection, or may define his own connection type as needed. Connections with other nodes in the social networking system 930, such as non-human entities, buckets, cluster centers, images, interests, pages, external systems, concepts, etc. are also stored in the connection store 938.

Social networking system 930 maintains data about objects with which the user can interact. In order to maintain this data, user profile store 936 and connection store 938 store corresponding type instances of objects maintained by social networking system 930. Each object type has an information field suitable for storing information suitable for the type of object. For example, the user profile store 936 includes a data structure having fields suitable for describing the user's account and information associated with the user's account. When a new object of a particular type is created, social networking system 930 initializes a new data structure of the corresponding type, assigns a unique object identifier to the new data structure, and In response, it begins to add data to the object. This occurs, for example, when a user becomes a user of social networking system 930, which creates a new instance of the user profile in user profile store 936; Assign a unique identifier to the user account and begin populating the user account fields with information provided by the user.

  Connection store 938 includes a data structure suitable for describing user connections to other users, connections to external systems 920, or connections to other entities. Connection store 938 may also associate connection types with user connections. Connection types can be used in conjunction with a user's privacy settings to regulate access to information about the user. In one embodiment of the present invention, user profile store 936 and connection store 938 may be implemented as a federated database.

  The data stored in the connection store 938, the user profile store 936, and the activity log 942 uses the nodes to identify the various objects and edges that connect the nodes and to establish relationships between the different objects. Allows social networking system 930 to generate a social graph to identify. For example, if a first user establishes a connection with a second user within the social networking system 930, the first user's and second user's user accounts from the user profile store 936 are Can act as a node in the social graph. The connection between the first user and the second user, stored by the connection store 938, is an edge between the nodes associated for the first user and the second user. Continuing with this example, the second user may then send a message within the social networking system 930 to the first user. The action of sending a message is another edge between two nodes in the social graph that represents the first user and the second user, and the action of sending a message can be stored. Also, the message itself can be identified in the social graph and included in the social graph as another node connected to the node representing the first user and the node representing the second user.

  In another example, the first user may be in an image maintained by social networking system 930 (or alternatively in an image maintained by another system external to social networking system 930). Two users may be tagged. The image may itself be represented as a node in the social networking system 930. This tagging action not only creates an edge between each of the users and the image, but can also create an edge between the first user and the second user. Images are also nodes in the social graph. In yet another example, if a user confirms that he / she will participate in an event, the user and the event are nodes obtained from the user profile store 936 and participation in the event is recorded in the activity log 942. Is an edge between nodes that can be obtained from By generating and maintaining social graphs, social networking system 930 provides socially relevant information, including data describing many different types of objects, as well as interactions and connections between those objects. Providing a rich source of.

Web server 932 links social networking system 930 through network 950 to one or more user devices 910 and / or one or more external systems 920. Web server 932 is a web page, and Java (registered trademark), JavaScript, Flash, XML
Supply other web-related content such as. Web server 932 includes a mail server, or may include other messaging functionality for receiving and routing messages between social networking system 930 and one or more user devices 910. . The message can be an instant message, a queue message (eg, email), a text and SMS message, or any other suitable messaging format.

  The API request server 934 allows one or more external systems 920 and user devices 910 to call access information from the social networking system 930 by calling one or more API functions. . The API request server 934 may also allow the external system 920 to send information to the social networking system 930 by calling the API. External system 920, in one embodiment, sends an API request to social networking system 930 through network 950, and API request server 934 receives the API request. The API request server 934 processes the request by calling the API associated with the API request and generates an appropriate response. The API request server 934 communicates the appropriate response to the external system 920 through the network 950. For example, in response to an API request, the API request server 934 collects data associated with the user, such as the connection of the user logged in to the external system 920, and the collected data to the external system 920 connect. In another embodiment, user device 910 communicates with social networking system 930 through an API in the same manner as external system 920.

  Action logger 940 may receive communications from web server 932 regarding user actions within social networking system 930 and / or outside social networking system 930. The action logger 940 populates the activity log 942 with information about user actions, and various actions performed by the user inside the social networking system 930 and outside the social networking system 930. Allows social networking system 930 to discover. Any action that a particular user takes with respect to another node on social networking system 930 can be determined by each user's account through information maintained in activity log 942 or a similar database or other data repository. Can be associated with. Examples of actions performed by a user within social networking system 930 that are identified and stored include, for example, adding a connection to another user, sending a message to another user, from another user Read messages, watch content associated with another user, participate in an event posted by another user, post an image, try to post an image, or otherwise Other actions that interact with other users or other objects. When a user performs an action in social networking system 930, the action is recorded in activity log 942. In one embodiment, social networking system 930 maintains activity log 942 as a database of entries. If an action is performed within the social networking system 930, an entry for the action is added to the activity log 942. The activity log 942 may be referred to as an action log.

User actions may also be associated with concepts and actions that occur within entities external to social networking system 930, such as external system 920 that is separate from social networking system 930. For example,
Action logger 940 may receive data from web server 932 describing the user's interaction with external system 920. In this example, external system 920 reports user interaction according to structured actions and objects in the social graph.

  Other examples of actions that a user interacts with external system 920 include a user expressing interest in external system 920 or another entity, a social discussion that discusses with external system 920 or web page 922a within external system 920. The user posts a comment to the networking system 930, and the user posts a uniform resource locator (URL) to the social networking system 930 or other identifier associated with the external system 920. , Participating in events associated with the external system 920, or any other action by the user related to the external system 920. Accordingly, activity log 942 may include actions that describe the interaction between a user of social networking system 930 and an external system 920 that is separate from social networking system 930.

  The authentication server 944 enhances one or more privacy settings of the user of the social networking system 930. A user's privacy settings determine how specific information associated with the user can be shared. A privacy setting comprises a specification of specific information associated with a user and a specification of one or more entities with which the information can be shared. Examples of entities with which information can be shared may include other users, applications, external systems 920, or any entity that can potentially access that information. Information that can be shared by the user is done by the user, such as user account information such as profile photos, phone numbers associated with the user, user connections, adding connections, changing user profile information, etc. Provide actions.

  Privacy setting specifications may be provided at different levels of granularity. For example, the privacy settings may identify specific information that should be shared with other users. The privacy setting identifies a specific set of related information, such as a work phone number or personal information including profile photo, home phone number, and status. Alternatively, privacy settings may be applied to all information associated with the user. The specification of the set of entities that can access specific information can also be specified at various levels of granularity. Various sets of entities with which information may be shared may include, for example, all friends of a user, all friends of a friend, all applications, or all external systems 920. One embodiment allows a specification of a set of entities to comprise a list of entities. For example, a user may provide a list of external systems 920 that are allowed to access certain information. Another embodiment allows a specification to include a set of entities with exceptions that are not allowed to access the information. For example, the user may specify a list of external systems 920 that may allow all external systems 920 to access the user's work information but are not allowed to access work information. Certain embodiments refer to a list of exceptions that are not allowed to access certain information as a “block list”. External systems 920 belonging to the block list specified by the user are blocked from accessing the information specified in the privacy settings. Various combinations of the granularity of the information specification and the granularity of the specification of the entity with which the information is shared are possible. For example, all personal information can be shared with friends, whereas all workplace information can be shared with friends of friends.

Authentication server 944 includes logic for determining whether certain information associated with the user can be accessed by the user's friends, external system 920, and / or other applications and entities. External system 920 may require authentication from authentication server 944 in order to access the user's more personal and sensitive information, such as the user's work phone number. Based on the user's privacy settings, the authentication server 944 may access information associated with the user, including information about actions taken by the user, by another user, external system 920, application, or another entity. Decide what you are allowed to do.

  Social networking system 930 may include a facial expression module 946. In some embodiments, the facial expression conversion module 946 may be implemented as the facial expression module 102 of FIG. Face representation conversion module 946 may be configured to align face images, classify face images, and / or verify face images by employing a deep neural network in accordance with embodiments of the present disclosure. .

Hardware Implementation The aforementioned processes and features may be implemented by a wide variety of machine and computer system architectures and in a wide variety of networks and computing environments. FIG. 10 illustrates an example computer system 1000 that can be used to implement one or more of the embodiments described herein, according to one embodiment of the invention. Computer system 1000 includes a set of instructions that cause computer system 1000 to perform the processes and features discussed herein. The computer system 1000 can be connected (eg, networked) to other machines. In a networked deployment, the computer system 1000 operates in the qualification of a server machine or client machine in a client / server network environment, or in a peer-to-peer (or distributed) network environment. Can act as a peer machine. In one embodiment of the present invention, computer system 1000 may be social networking system 1030, user device 910, and external system 1020, or components thereof. In one embodiment of the present invention, the computer system 1000 may be one of many servers that make up all or part of the social networking system 1030.

Computer system 1000 includes a processor 1002, a cache 1004, one or more executable modules and drivers stored on a computer-readable medium and directed to the processes and features described herein. Including. The computer system 1000 also includes a high performance input / output (I / O) bus 1006 and a standard I / O bus 1008. Host bridge 1010 couples processor 1002 to high performance I / O bus 1006, while I / O bus bridge 1012 couples two buses 1006 and 1008 together. System memory 1014 and one or more network interfaces 1016 are coupled to a high performance I / O bus 1006. Computer system 1000 may further include a video memory and a display device coupled to the video memory (not shown). Mass storage 1018 and I / O port 1020 are coupled to standard I / O bus 1008. Computer system 1000 may optionally include a keyboard and pointing device, display device, or other input / output device (not shown) coupled to standard I / O bus 1008. Overall, these elements are x86 compatible processors manufactured by Intel Corporation of Santa Clara, California, Advanced Micro Devices (AMD), Sunnyvale, California. It is intended to represent a broad category of computer hardware systems including, but not limited to, x86 compatible processors manufactured by Devices, Inc), and any other suitable processor. .

  The operating system manages and controls the operation of the computer system 1000, including the input of data to a software application (not shown) and the output of data from the software application. The operating system provides an interface between software applications running on the system and the hardware components of the system. Any suitable operating system, such as the LINUX operating system, the Apple Macintosh operating system commercially available from Apple Computer, Inc., Cupertino, California, UNIX operating system A system, a Microsoft® Windows® operating system, a BSD operating system, etc. may be used. Other implementations are possible.

  The elements of computer system 1000 are described in more detail below. Specifically, the network interface 1016 provides communication between the computer system 1000 and any of a wide range of networks, such as an Ethernet (eg, IEEE 802.3) network, backplane, etc. To do. Mass storage 1018 provides persistent storage for data and programming instructions for executing the above-described processes and features implemented by each of the above-identified computing systems. The system memory 1014 (eg, DRAM) provides temporary storage for data and programming instructions when executed by the processor 1002. I / O port 1020 may be one or more serial and / or parallel communication ports that provide communication between additional peripheral devices that may be coupled to computer system 1000.

  Computer system 1000 can include a variety of system architectures, and the various components of computer system 1000 can be rearranged. For example, the cache 1004 may be implemented on a chip with the processor 1002. Alternatively, cache 1004 and processor 1002 may be grouped together as a “processor module” and processor 1002 may be referred to as a “processor core”. Further, certain embodiments of the invention may not require or include all of the above components. For example, a peripheral device coupled to the standard I / O bus 1008 may be coupled to the high performance I / O bus 1006. Also, in some embodiments, there may be only a single bus and the components of computer system 1000 may be coupled to that single bus. Further, the computer system 1000 may include additional components such as additional processors, storage devices, or memories.

  In general, the processes and features described herein may be implemented as part of an operating system or a particular application, component, program, object, module, or series of instructions referred to as a “program”. For example, one or more programs may be used to perform the specific processes described herein. A program, when read and executed by one or more processors, typically acts on computer system 1000 to perform the processes and features described herein. One or more instructions to be executed are provided in various memory and storage devices in computer system 1000. The processes and features described herein may be implemented in software, firmware, hardware (eg, application specific integrated circuits), or any combination thereof.

  In one implementation, the processes and features described herein are implemented as a series of executable modules that are executed by computer system 1000 either individually or collectively in a distributed computing environment. The aforementioned modules may be implemented by hardware, executable modules stored on computer readable media (or machine readable media), or a combination of both. For example, a module may comprise multiple instructions or a series of instructions to be executed by a processor in a hardware system, such as processor 1002. First, the sequence of instructions may be stored on a storage device such as mass storage 1018. However, the series of instructions may be stored on any suitable computer readable storage medium. Further, the sequence of instructions need not be stored locally, but may be received through the network interface 1016 from a remote storage device such as a server on the network. The instructions are copied from a storage device, such as mass storage 1018, into system memory 1014 and then accessed and executed by processor 1002. In various embodiments, one or more modules may be executed by one or more processors at one or more locations, such as multiple servers in a parallel processing environment.

  Examples of computer readable media include recordable type media such as volatile and non-volatile memory devices, solid state memory, floppy disks and other removable disks, hardware Disk drives, magnetic media, optical disks (eg, compact disk read only memory (CD ROMS), digital versatile disk (DVD)), other similar non-temporary (or temporary), tangible ( (Or intangible) storage medium or a sequence of instructions for execution by the computer system 1000 to execute any one or more of the processes and features described herein, Including any type of media suitable for crystallization or support, But it is not limited.

  For the purpose of explanation, numerous specific details are set forth in order to provide a thorough understanding of the present description. However, it will be apparent to those skilled in the art that embodiments of the present disclosure may be practiced without these specific details. In some instances, modules, structures, processes, features, and devices are shown in block diagram form in order to avoid obscuring the present description. In other examples, functional block diagrams and flow diagrams are shown to represent data and logic flows. Block diagram and flow diagram components (eg, modules, blocks, structures, devices, features, etc.) may be variously combined and separated in ways other than those explicitly described and illustrated herein. May be removed, rearranged and replaced.

  References to “one embodiment”, “one embodiment”, “other embodiments”, “a series of embodiments”, “some embodiments”, “various embodiments”, etc. in this specification. Means that a particular feature, design, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the present disclosure. The appearances of the phrases “in one embodiment” or “in one embodiment” in various parts of the specification are not necessarily all referring to the same embodiment, but separate embodiments or alternatives. Such embodiments are not mutually exclusive with other embodiments. Furthermore, various features are described, with or without explicit reference to “embodiments” and the like, and various features can be variously combined and included in some embodiments. However, various other embodiments may be omitted. Similarly, various features are described which may be preferences or requirements for some embodiments but not preferences or requirements for other embodiments.

The terminology used herein is selected primarily for readability and educational purposes and is not selected to outline or delineate the subject matter of the invention. Therefore, it is intended that the scope of the invention be limited not by this detailed description, but rather by any claim issued for an application based on this specification. Accordingly, the disclosure of embodiments of the invention is intended to be illustrative of the scope of the invention, but not to be limiting, the scope of the invention being set forth in the following claims.

Claims (19)

A computing system determining a face portion of the two-dimensional image based at least in part on the first support vector regressor trained to predict the point configuration;
Said computing system determining a set of reference points of a two-dimensional face image generated by transforming a face portion of said two-dimensional image, wherein at least some reference points are defined as facial features. Correspondingly, the set of reference points is localized based at least in part on a second support vector regressor that is different from the first support vector regressor;
The computing system is based at least in part on the set of reference points detected in the two-dimensional face image and a corresponding set of anchor points arranged in a three-dimensional shape of a human face. a step of generating a three-dimensional aligned face images from the 2-dimensional face image,
Before SL computing system, the method comprising, based on the provision of the 3-dimensional alignment face image to deep neural network (DNN), to classify the face ID corresponding to the 2-dimensional face image, The DNN is trained to classify the two-dimensional face image based on the set of features of the three-dimensionally aligned face image . The step of generating a three-dimensionally aligned face image includes :
A step of detecting a set of reference points prior SL 2-dimensional face image,
Using a set of pre-Symbol reference point, and a step of processing the 2-dimensional face images to 3-dimensional shape, the computer implemented method of claim 1. As engineering for the processing,
A set of anchors point comprising the steps of placing into said three-dimensional shape on each anchor point corresponds to one reference point of the set of the reference point of the 2-dimensional face image, placing Process ,
A step of back projection position of each reference point of the set of the reference point before Symbol 2-dimensional face images to said three-dimensional shape,
A step of projecting the position of each reference point of the set of the reference point before Symbol 2-dimensional face image to the front image of the three-dimensional shape,
Based on a set of a set of pre-Symbol reference point the anchor point, and a step of converting the 2-dimensional face image to the three-dimensional alignment face image, a computer implemented method of claim 2. Step set of reference points of the 2-dimensional face images is to define a set of triangles, the step of converting into the three-dimensional alignment face image is, to perform the affine transformation of each triangle of the set of the triangle The computer-implemented method of claim 3, comprising: And identifying by the face portion of the images, it detects a second set of reference points of the image,
Generating the two-dimensional face image by transforming the face portion of the image based on the second set of reference points .
The computer-implemented method of claim 1 . 6. The computer of claim 5, wherein generating the two-dimensional face image comprises determining a set of anchor positions and processing the face portion of the image based on the set of anchor positions. Implementation method. The computer-implemented method of claim 1 , wherein the DNN comprises a set of layers, the layers comprising a convolutional pooling layer, a set of local coupling layers, and a set of full coupling layers. The convolution pooling layer comprises a first convolution layer, a maximum pooling layer, and a second convolution layer, wherein the convolution pooling layer extracts a set of features of the facial image that are three-dimensionally aligned. configured the computer-implemented method of claim 7. The computer-implemented method of claim 7, wherein each local coupling layer of the set of local coupling layers is configured to extract a set of features of the facial image that are three-dimensionally aligned. The computer-implemented method of claim 7, wherein each full connected layer of the set of full connected layers is configured to determine a correlation within the set of facial image features that are three-dimensionally aligned. The computer-implemented method of claim 7, wherein an output of one fully connected layer of the set of fully connected layers is a feature vector. The computer-implemented method of claim 7, wherein each feature of the set of features is normalized to a predetermined range. The computer-implemented method of claim 1 , wherein the DNN is configured to define a set of filters for each pixel of the face image that is three-dimensionally aligned. The computer-implemented method of claim 13 , further comprising training the DNN to define the set of filters using a data set. Further comprising identifying an individual of the two-dimensional face image by querying the ID of the two-dimensional face image in an image store, the image store comprising a set of face images, of the set of face images each face image corresponds to the personal computer-implemented method of claim 1. A step of classifying the second ID of the second 2-dimensional face image,
For prior Symbol ID and the second ID to determine whether corresponding to individual, further comprising the step of comparing the ID for the second ID, computer implementation according to claim 1 Method. The ID comprises a first set of features, the second ID comprises a second set of features, and the step of comparing IDs comprises :
Determining a set of characteristic differences between the second set of the first set and features of Features,
17. The computer-implemented method of claim 16 , comprising: mapping the set of feature differences to a predetermined value by using a fully connected layer. At least one processor;
And a memory for storing instructions, wherein when the instructions are executed by the at least one processor, the system includes :
Determining a facial portion of the two-dimensional image based at least in part on the first support vector regressor trained to predict the point composition;
Determining a set of reference points of a two-dimensional face image generated by transforming a face portion of the two-dimensional image, wherein at least some reference points correspond to facial features, The set of reference points is localized based at least in part on a second support vector regressor different from the support vector regressor;
From the two-dimensional face image based at least in part on the set of reference points detected in the two-dimensional face image and a set of corresponding anchor points arranged in a three-dimensional shape of a human face. Generating a three-dimensionally aligned face image ;
Based on providing the 3-dimensional alignment face image to de Ipu neural network (DNN), a step of classifying the face ID corresponding to the 2-dimensional face image, the DNN, the 3-dimensional A system trained to classify the two-dimensional face image based on a set of features of the aligned face image . A non-transitory computer readable storage medium containing instructions, wherein the instructions are executed by the computing system when executed by at least one processor of the computing system .
Determining a facial portion of the two-dimensional image based at least in part on the first support vector regressor trained to predict the point composition;
Determining a set of reference points of a two-dimensional face image generated by transforming a face portion of the two-dimensional image, wherein at least some reference points correspond to facial features, The set of reference points is localized based at least in part on a second support vector regressor different from the support vector regressor;
From the two-dimensional face image based at least in part on the set of reference points detected in the two-dimensional face image and a set of corresponding anchor points arranged in a three-dimensional shape of a human face. Generating a three-dimensionally aligned face image;
Based on providing the 3-dimensional alignment face image to de Ipu neural network (DNN), a step of classifying the face ID corresponding to the 2-dimensional face image, the DNN, the 3-dimensional A storage medium that is trained to classify the two-dimensional face image based on a set of features of the aligned face image .